Trustworthy Scientific Literature Discovery Using Knowledge Graphs and Multi-Role Large Language Model Reasoning

Abstract

Scientific literature discovery has become a critical infrastructure problem as publication volumes, heterogeneous evidence forms, and interdisciplinary dependencies overwhelm conventional search and recommendation systems. This paper examines how knowledge graphs and multi-role large language model reasoning can be combined to support trustworthy discovery. The argument begins from the premise that trustworthiness is not a property of model accuracy alone but of the entire socio-technical pipeline through which claims, metadata, provenance, rankings, and explanations are produced. Knowledge graphs provide structured representations of papers, authors, methods, datasets, venues, citations, and epistemic relations, while large language models can interpret ambiguous queries, synthesize evidence, critique methodological fit, and rerank candidate literature according to multiple evaluative perspectives. A multi-role architecture is proposed in conceptual terms, with roles such as query interpreter, evidence retriever, methodological critic, domain synthesizer, adversarial reviewer, and compliance auditor. The paper analyzes architectural trade-offs involving latency, cost, freshness, explainability, governance, fairness, sustainability, and robustness. It further discusses deployment environments, audit mechanisms, model and dataset documentation, hallucination mitigation, and policy implications for research evaluation. The central conclusion is that trustworthy discovery requires a layered system in which knowledge graphs constrain generative reasoning, multi-role orchestration exposes disagreement, and governance mechanisms preserve accountability across institutions and domains.